FADA: Accessible fetal ultrasound interpretation and annotation with a selectively distilled unified vision-language model

A single vision-language model called FADA can perform clinical interpretation, classification, detection, and segmentation of fetal ultrasound images without requiring external labels, according to research published on arXiv. The system is designed to operate offline on a commodity smartphone, targeting diagnostic access gaps in settings where trained sonographers are scarce [1][2]. The model, built on the Qwen3.5-VL architecture, distills knowledge from four domain-specific foundation models — FetalCLIP, UltraSAM, USF-MAE, and UltraFedFM — using offline pre-computed feature caching [1][2]. A technique termed selective distillation, which applies feature alignment only to annotation tasks while interpretation relies on standard fine-tuning, consistently outperformed full distillation across most evaluation metrics [2]. The recommended variant, FADA-SKD, achieved a 0.8820 mean Dice score for segmentation and 0.7671 [email protected] for detection [1][2]. Structured interpretation compliance reached 100 percent [2]. Expert sonographer validation across 237 images confirmed clinically acceptable outputs in both autonomous and human-in-the-loop modes, with 73.5 percent of interpretations scoring perfectly under clinician guidance [1][2]. The system is trainable on a single consumer GPU and does not require cloud connectivity for deployment [1][2]. Researchers validated edge deployment by running a compressed 0.8-billion-parameter version of the model on a Qualcomm Snapdragon 7 Gen 1 smartphone with 12 GB of RAM, using llama.cpp with GGUF quantization. The full five-phase pipeline completed in approximately 60 seconds entirely offline [1][2]. The work directly addresses a persistent global health challenge. A shortage of trained sonographers limits prenatal ultrasound screening in low- and middle-income countries, where over half of pregnant women receive no skilled sonography [1][2]. Current deep learning approaches typically address detection, segmentation, or classification in isolation, each requiring a separate model and expert-specified labels at inference [2]. FADA consolidates these tasks into a single interpretation-first pipeline, establishing a practical pathway for integrating AI-assisted fetal assessment with portable ultrasound devices [1][2]. The United Nations Sustainable Development Goals, adopted in 2015, include good health and well-being as a core target, though progress has been uneven and funding remains a critical issue [6]. Code, models, and data for FADA have been made publicly available on GitHub by lead researcher Mahmood Saleh Alzubaidi [1][2].